Demis Hassabis, CEO of Google DeepMind, predicts that in the near future, artificial intelligence systems will no longer be limited to answering questions, but will be able to autonomously plan and take action.

In an interview, Hassabis revealed that they are developing such "agent-based" systems and are expected to be deployed in the next one to two years.

"I am very excited about the next stage of these large-scale general models. I believe we will see more behavior similar to that of agents, possibly as early as this year or next year," said Hassabis.

He further pointed out that these systems can not only answer questions but also autonomously plan and take action in the real world. Hassabis believes that this ability to set and achieve goals will make these systems more practical as everyday tools.

Hassabis firmly believes that this is a crucial step in making artificial intelligence systems more useful in everyday life. DeepMind can draw on its rich experience in reinforcement learning, a technology that has been applied in the famous Go program AlphaGo.

"We are working on it, and so are others. It reminds us of the work we did in the gaming field a few years ago, which were all agent systems dedicated to achieving certain goals and tasks. We are now combining this work with the current large-scale multimodal work," explained Hassabis.

Action and Vision

Prior to the release of the Gemini language model in June 2023, Hassabis hinted in an interview with Wired magazine that DeepMind's expertise in reinforcement learning could give Gemini a unique advantage.

With the help of AlphaGo's reinforcement learning and tree search techniques, Gemini is expected to demonstrate outstanding problem-solving and planning abilities in the future.

Hassabis's comments indicate that DeepMind is steadily advancing the development of its AI systems to achieve higher autonomy.

If DeepMind successfully develops AI agents capable of independently solving complex tasks in the real world, their impact will be profound. These systems have a wide range of potential applications, including intelligent personal assistants, autonomous robots, and self-learning systems in scientific research.

In addition to Gemini, DeepMind is also working on projects such as the RT model, which uses large-scale AI models for image and language processing to give AI embedded in robots greater action capabilities in daily life.

DeepMind's approach of combining language models with agent technology is similar to that of companies like OpenAI and Anthropic, which are also actively developing AI agents capable of communication and action.

OpenAI recently re-entered the field of robotics, combining its vision language model with Figure's robot technology.

Hassabis firmly believes that AI will benefit humanity

Despite critics warning that large-scale AI models require massive energy support during training and inference, and that energy requirements continue to increase with the scale of the models, Hassabis disagrees and firmly believes that the massive resources invested in AI will ultimately bring substantial returns to humanity.

"In the long run, I believe the benefits brought by generative AI models we build in fields such as drug discovery will far outweigh these costs," said Hassabis.

He also sees the potential value of AI in areas such as energy and climate, such as through more efficient power grids, new materials, and technologies. He believes that these technologies could be highly efficient and useful, far exceeding the costs and efforts required to build these systems.

Ultimately, Hassabis stated that their goal is to apply AI in as many fields as possible. Only then can the development of AI be sustainable.